We trained a large, deep convolutional neural network to classify the 1.2 million high-resolution images in the ImageNet LSVRC-2010 contest into the 1000 different classes. On the test data, we achieved top-1 and top-5 error rates of 37.5% and 17.0% which is considerably better than the previous state-of-the-art. The neural network, which has 60 million parameters and 650,000 neurons, consists of five convolutional layers, some of which are followed by max-pooling layers, and three fully-connected layers with a final 1000-way softmax. To make training faster, we used non-saturating neurons and a very efficient GPU implementation of the convolution operation. To reduce overfitting in the fully-connected layers we employed a recently-developed regularization method called “dropout” that proved to be very effective. We also entered a variant of this model in the ILSVRC-2012 competition and achieved a winning top-5 test error rate of 15.3%, compared to 26.2% achieved by the second-best entry.
核心贡献 · Key contributions
在 ImageNet 上训练了一个 60M 参数的大型深度 CNN，在 ILSVRC-2010 上取得 17.0%的 top-5 错误率。 Trained a large deep CNN on ImageNet with 60M parameters, achieving top-5 error of 17.0% on ILSVRC-2010.
引入 ReLU 非线性加速训练，超越 tanh 等饱和激活函数。 Introduced ReLU nonlinearity to accelerate training, surpassing saturating activations like tanh.
使用 dropout 正则化有效减少大型全连接层中的过拟合。 Used dropout regularization to effectively reduce overfitting in large fully connected layers.
证明了深度的重要性：移除一个卷积层会使 top-1 准确率下降约 2%。 Demonstrated depth importance: removing a convolutional layer degrades top-1 accuracy by about 2%.
局限 · Limitations
网络规模受限于 GPU 内存（3GB），需要多 GPU 和数天训练时间。 Network size limited by GPU memory (3GB); requires multiple GPUs and days of training.
模型严重依赖大规模标注数据集；数据不足时性能下降。 Model heavily reliant on large labeled datasets; performance drops without sufficient data.
未使用无监督预训练；利用额外无标签数据可能进一步提升性能。 No unsupervised pre-training used; may improve with additional unlabeled data.
架构针对 ImageNet 优化；向其他任务的迁移性未充分验证。 Architecture optimized for ImageNet; transferability to other tasks not extensively validated.
Dropout 使训练迭代次数加倍；尽管优化，计算成本仍然较高。 Dropout doubles training iterations; computational cost still high despite optimizations.